1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 * Copyright (c) 2009, Microsoft Corporation.
   4 *
 
 
 
 
 
 
 
 
 
 
 
 
   5 * Authors:
   6 *   Haiyang Zhang <haiyangz@microsoft.com>
   7 *   Hank Janssen  <hjanssen@microsoft.com>
   8 */
   9#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  10
  11#include <linux/kernel.h>
  12#include <linux/sched.h>
  13#include <linux/wait.h>
  14#include <linux/mm.h>
  15#include <linux/delay.h>
  16#include <linux/io.h>
  17#include <linux/slab.h>
  18#include <linux/netdevice.h>
  19#include <linux/if_ether.h>
  20#include <linux/vmalloc.h>
  21#include <linux/rtnetlink.h>
  22#include <linux/prefetch.h>
  23#include <linux/filter.h>
  24
  25#include <asm/sync_bitops.h>
  26#include <asm/mshyperv.h>
  27
  28#include "hyperv_net.h"
  29#include "netvsc_trace.h"
  30
  31/*
  32 * Switch the data path from the synthetic interface to the VF
  33 * interface.
  34 */
  35int netvsc_switch_datapath(struct net_device *ndev, bool vf)
  36{
  37	struct net_device_context *net_device_ctx = netdev_priv(ndev);
  38	struct hv_device *dev = net_device_ctx->device_ctx;
  39	struct netvsc_device *nv_dev = rtnl_dereference(net_device_ctx->nvdev);
  40	struct nvsp_message *init_pkt = &nv_dev->channel_init_pkt;
  41	int ret, retry = 0;
  42
  43	/* Block sending traffic to VF if it's about to be gone */
  44	if (!vf)
  45		net_device_ctx->data_path_is_vf = vf;
  46
  47	memset(init_pkt, 0, sizeof(struct nvsp_message));
  48	init_pkt->hdr.msg_type = NVSP_MSG4_TYPE_SWITCH_DATA_PATH;
  49	if (vf)
  50		init_pkt->msg.v4_msg.active_dp.active_datapath =
  51			NVSP_DATAPATH_VF;
  52	else
  53		init_pkt->msg.v4_msg.active_dp.active_datapath =
  54			NVSP_DATAPATH_SYNTHETIC;
  55
  56again:
  57	trace_nvsp_send(ndev, init_pkt);
  58
  59	ret = vmbus_sendpacket(dev->channel, init_pkt,
  60			       sizeof(struct nvsp_message),
  61			       (unsigned long)init_pkt, VM_PKT_DATA_INBAND,
  62			       VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
  63
  64	/* If failed to switch to/from VF, let data_path_is_vf stay false,
  65	 * so we use synthetic path to send data.
  66	 */
  67	if (ret) {
  68		if (ret != -EAGAIN) {
  69			netdev_err(ndev,
  70				   "Unable to send sw datapath msg, err: %d\n",
  71				   ret);
  72			return ret;
  73		}
  74
  75		if (retry++ < RETRY_MAX) {
  76			usleep_range(RETRY_US_LO, RETRY_US_HI);
  77			goto again;
  78		} else {
  79			netdev_err(
  80				ndev,
  81				"Retry failed to send sw datapath msg, err: %d\n",
  82				ret);
  83			return ret;
  84		}
  85	}
  86
  87	wait_for_completion(&nv_dev->channel_init_wait);
  88	net_device_ctx->data_path_is_vf = vf;
  89
  90	return 0;
  91}
  92
  93/* Worker to setup sub channels on initial setup
  94 * Initial hotplug event occurs in softirq context
  95 * and can't wait for channels.
  96 */
  97static void netvsc_subchan_work(struct work_struct *w)
  98{
  99	struct netvsc_device *nvdev =
 100		container_of(w, struct netvsc_device, subchan_work);
 101	struct rndis_device *rdev;
 102	int i, ret;
 103
 104	/* Avoid deadlock with device removal already under RTNL */
 105	if (!rtnl_trylock()) {
 106		schedule_work(w);
 107		return;
 108	}
 109
 110	rdev = nvdev->extension;
 111	if (rdev) {
 112		ret = rndis_set_subchannel(rdev->ndev, nvdev, NULL);
 113		if (ret == 0) {
 114			netif_device_attach(rdev->ndev);
 115		} else {
 116			/* fallback to only primary channel */
 117			for (i = 1; i < nvdev->num_chn; i++)
 118				netif_napi_del(&nvdev->chan_table[i].napi);
 119
 120			nvdev->max_chn = 1;
 121			nvdev->num_chn = 1;
 122		}
 123	}
 124
 125	rtnl_unlock();
 126}
 127
 128static struct netvsc_device *alloc_net_device(void)
 129{
 130	struct netvsc_device *net_device;
 131
 132	net_device = kzalloc(sizeof(struct netvsc_device), GFP_KERNEL);
 133	if (!net_device)
 134		return NULL;
 135
 136	init_waitqueue_head(&net_device->wait_drain);
 137	net_device->destroy = false;
 138	net_device->tx_disable = true;
 139
 140	net_device->max_pkt = RNDIS_MAX_PKT_DEFAULT;
 141	net_device->pkt_align = RNDIS_PKT_ALIGN_DEFAULT;
 142
 143	init_completion(&net_device->channel_init_wait);
 144	init_waitqueue_head(&net_device->subchan_open);
 145	INIT_WORK(&net_device->subchan_work, netvsc_subchan_work);
 146
 147	return net_device;
 148}
 149
 150static void free_netvsc_device(struct rcu_head *head)
 151{
 152	struct netvsc_device *nvdev
 153		= container_of(head, struct netvsc_device, rcu);
 154	int i;
 155
 156	kfree(nvdev->extension);
 
 
 
 157
 158	if (!nvdev->recv_buf_gpadl_handle.decrypted)
 159		vfree(nvdev->recv_buf);
 160	if (!nvdev->send_buf_gpadl_handle.decrypted)
 161		vfree(nvdev->send_buf);
 162	bitmap_free(nvdev->send_section_map);
 163
 164	for (i = 0; i < VRSS_CHANNEL_MAX; i++) {
 165		xdp_rxq_info_unreg(&nvdev->chan_table[i].xdp_rxq);
 166		kfree(nvdev->chan_table[i].recv_buf);
 167		vfree(nvdev->chan_table[i].mrc.slots);
 168	}
 169
 170	kfree(nvdev);
 171}
 172
 173static void free_netvsc_device_rcu(struct netvsc_device *nvdev)
 174{
 175	call_rcu(&nvdev->rcu, free_netvsc_device);
 176}
 177
 178static void netvsc_revoke_recv_buf(struct hv_device *device,
 179				   struct netvsc_device *net_device,
 180				   struct net_device *ndev)
 181{
 182	struct nvsp_message *revoke_packet;
 183	int ret;
 184
 185	/*
 186	 * If we got a section count, it means we received a
 187	 * SendReceiveBufferComplete msg (ie sent
 188	 * NvspMessage1TypeSendReceiveBuffer msg) therefore, we need
 189	 * to send a revoke msg here
 190	 */
 191	if (net_device->recv_section_cnt) {
 192		/* Send the revoke receive buffer */
 193		revoke_packet = &net_device->revoke_packet;
 194		memset(revoke_packet, 0, sizeof(struct nvsp_message));
 195
 196		revoke_packet->hdr.msg_type =
 197			NVSP_MSG1_TYPE_REVOKE_RECV_BUF;
 198		revoke_packet->msg.v1_msg.
 199		revoke_recv_buf.id = NETVSC_RECEIVE_BUFFER_ID;
 200
 201		trace_nvsp_send(ndev, revoke_packet);
 202
 203		ret = vmbus_sendpacket(device->channel,
 204				       revoke_packet,
 205				       sizeof(struct nvsp_message),
 206				       VMBUS_RQST_ID_NO_RESPONSE,
 207				       VM_PKT_DATA_INBAND, 0);
 208		/* If the failure is because the channel is rescinded;
 209		 * ignore the failure since we cannot send on a rescinded
 210		 * channel. This would allow us to properly cleanup
 211		 * even when the channel is rescinded.
 212		 */
 213		if (device->channel->rescind)
 214			ret = 0;
 215		/*
 216		 * If we failed here, we might as well return and
 217		 * have a leak rather than continue and a bugchk
 218		 */
 219		if (ret != 0) {
 220			netdev_err(ndev, "unable to send "
 221				"revoke receive buffer to netvsp\n");
 222			return;
 223		}
 224		net_device->recv_section_cnt = 0;
 225	}
 226}
 227
 228static void netvsc_revoke_send_buf(struct hv_device *device,
 229				   struct netvsc_device *net_device,
 230				   struct net_device *ndev)
 231{
 232	struct nvsp_message *revoke_packet;
 233	int ret;
 234
 235	/* Deal with the send buffer we may have setup.
 236	 * If we got a  send section size, it means we received a
 237	 * NVSP_MSG1_TYPE_SEND_SEND_BUF_COMPLETE msg (ie sent
 238	 * NVSP_MSG1_TYPE_SEND_SEND_BUF msg) therefore, we need
 239	 * to send a revoke msg here
 240	 */
 241	if (net_device->send_section_cnt) {
 242		/* Send the revoke receive buffer */
 243		revoke_packet = &net_device->revoke_packet;
 244		memset(revoke_packet, 0, sizeof(struct nvsp_message));
 245
 246		revoke_packet->hdr.msg_type =
 247			NVSP_MSG1_TYPE_REVOKE_SEND_BUF;
 248		revoke_packet->msg.v1_msg.revoke_send_buf.id =
 249			NETVSC_SEND_BUFFER_ID;
 250
 251		trace_nvsp_send(ndev, revoke_packet);
 252
 253		ret = vmbus_sendpacket(device->channel,
 254				       revoke_packet,
 255				       sizeof(struct nvsp_message),
 256				       VMBUS_RQST_ID_NO_RESPONSE,
 257				       VM_PKT_DATA_INBAND, 0);
 258
 259		/* If the failure is because the channel is rescinded;
 260		 * ignore the failure since we cannot send on a rescinded
 261		 * channel. This would allow us to properly cleanup
 262		 * even when the channel is rescinded.
 263		 */
 264		if (device->channel->rescind)
 265			ret = 0;
 266
 267		/* If we failed here, we might as well return and
 268		 * have a leak rather than continue and a bugchk
 269		 */
 270		if (ret != 0) {
 271			netdev_err(ndev, "unable to send "
 272				   "revoke send buffer to netvsp\n");
 273			return;
 274		}
 275		net_device->send_section_cnt = 0;
 276	}
 277}
 278
 279static void netvsc_teardown_recv_gpadl(struct hv_device *device,
 280				       struct netvsc_device *net_device,
 281				       struct net_device *ndev)
 282{
 283	int ret;
 284
 285	if (net_device->recv_buf_gpadl_handle.gpadl_handle) {
 286		ret = vmbus_teardown_gpadl(device->channel,
 287					   &net_device->recv_buf_gpadl_handle);
 288
 289		/* If we failed here, we might as well return and have a leak
 290		 * rather than continue and a bugchk
 291		 */
 292		if (ret != 0) {
 293			netdev_err(ndev,
 294				   "unable to teardown receive buffer's gpadl\n");
 295			return;
 296		}
 
 297	}
 298}
 299
 300static void netvsc_teardown_send_gpadl(struct hv_device *device,
 301				       struct netvsc_device *net_device,
 302				       struct net_device *ndev)
 303{
 304	int ret;
 305
 306	if (net_device->send_buf_gpadl_handle.gpadl_handle) {
 307		ret = vmbus_teardown_gpadl(device->channel,
 308					   &net_device->send_buf_gpadl_handle);
 309
 310		/* If we failed here, we might as well return and have a leak
 311		 * rather than continue and a bugchk
 312		 */
 313		if (ret != 0) {
 314			netdev_err(ndev,
 315				   "unable to teardown send buffer's gpadl\n");
 316			return;
 317		}
 
 318	}
 319}
 320
 321int netvsc_alloc_recv_comp_ring(struct netvsc_device *net_device, u32 q_idx)
 322{
 323	struct netvsc_channel *nvchan = &net_device->chan_table[q_idx];
 324	int node = cpu_to_node(nvchan->channel->target_cpu);
 325	size_t size;
 326
 327	size = net_device->recv_completion_cnt * sizeof(struct recv_comp_data);
 328	nvchan->mrc.slots = vzalloc_node(size, node);
 329	if (!nvchan->mrc.slots)
 330		nvchan->mrc.slots = vzalloc(size);
 331
 332	return nvchan->mrc.slots ? 0 : -ENOMEM;
 333}
 334
 335static int netvsc_init_buf(struct hv_device *device,
 336			   struct netvsc_device *net_device,
 337			   const struct netvsc_device_info *device_info)
 338{
 339	struct nvsp_1_message_send_receive_buffer_complete *resp;
 340	struct net_device *ndev = hv_get_drvdata(device);
 341	struct nvsp_message *init_packet;
 342	unsigned int buf_size;
 343	int i, ret = 0;
 
 344
 345	/* Get receive buffer area. */
 346	buf_size = device_info->recv_sections * device_info->recv_section_size;
 347	buf_size = roundup(buf_size, PAGE_SIZE);
 348
 349	/* Legacy hosts only allow smaller receive buffer */
 350	if (net_device->nvsp_version <= NVSP_PROTOCOL_VERSION_2)
 351		buf_size = min_t(unsigned int, buf_size,
 352				 NETVSC_RECEIVE_BUFFER_SIZE_LEGACY);
 353
 354	net_device->recv_buf = vzalloc(buf_size);
 355	if (!net_device->recv_buf) {
 356		netdev_err(ndev,
 357			   "unable to allocate receive buffer of size %u\n",
 358			   buf_size);
 359		ret = -ENOMEM;
 360		goto cleanup;
 361	}
 362
 363	net_device->recv_buf_size = buf_size;
 364
 365	/*
 366	 * Establish the gpadl handle for this buffer on this
 367	 * channel.  Note: This call uses the vmbus connection rather
 368	 * than the channel to establish the gpadl handle.
 369	 */
 370	ret = vmbus_establish_gpadl(device->channel, net_device->recv_buf,
 371				    buf_size,
 372				    &net_device->recv_buf_gpadl_handle);
 373	if (ret != 0) {
 374		netdev_err(ndev,
 375			"unable to establish receive buffer's gpadl\n");
 376		goto cleanup;
 377	}
 378
 379	/* Notify the NetVsp of the gpadl handle */
 380	init_packet = &net_device->channel_init_pkt;
 381	memset(init_packet, 0, sizeof(struct nvsp_message));
 382	init_packet->hdr.msg_type = NVSP_MSG1_TYPE_SEND_RECV_BUF;
 383	init_packet->msg.v1_msg.send_recv_buf.
 384		gpadl_handle = net_device->recv_buf_gpadl_handle.gpadl_handle;
 385	init_packet->msg.v1_msg.
 386		send_recv_buf.id = NETVSC_RECEIVE_BUFFER_ID;
 387
 388	trace_nvsp_send(ndev, init_packet);
 389
 390	/* Send the gpadl notification request */
 391	ret = vmbus_sendpacket(device->channel, init_packet,
 392			       sizeof(struct nvsp_message),
 393			       (unsigned long)init_packet,
 394			       VM_PKT_DATA_INBAND,
 395			       VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
 396	if (ret != 0) {
 397		netdev_err(ndev,
 398			"unable to send receive buffer's gpadl to netvsp\n");
 399		goto cleanup;
 400	}
 401
 402	wait_for_completion(&net_device->channel_init_wait);
 403
 404	/* Check the response */
 405	resp = &init_packet->msg.v1_msg.send_recv_buf_complete;
 406	if (resp->status != NVSP_STAT_SUCCESS) {
 407		netdev_err(ndev,
 408			   "Unable to complete receive buffer initialization with NetVsp - status %d\n",
 409			   resp->status);
 410		ret = -EINVAL;
 411		goto cleanup;
 412	}
 413
 414	/* Parse the response */
 415	netdev_dbg(ndev, "Receive sections: %u sub_allocs: size %u count: %u\n",
 416		   resp->num_sections, resp->sections[0].sub_alloc_size,
 417		   resp->sections[0].num_sub_allocs);
 418
 419	/* There should only be one section for the entire receive buffer */
 420	if (resp->num_sections != 1 || resp->sections[0].offset != 0) {
 421		ret = -EINVAL;
 422		goto cleanup;
 423	}
 424
 425	net_device->recv_section_size = resp->sections[0].sub_alloc_size;
 426	net_device->recv_section_cnt = resp->sections[0].num_sub_allocs;
 427
 428	/* Ensure buffer will not overflow */
 429	if (net_device->recv_section_size < NETVSC_MTU_MIN || (u64)net_device->recv_section_size *
 430	    (u64)net_device->recv_section_cnt > (u64)buf_size) {
 431		netdev_err(ndev, "invalid recv_section_size %u\n",
 432			   net_device->recv_section_size);
 433		ret = -EINVAL;
 434		goto cleanup;
 435	}
 436
 437	for (i = 0; i < VRSS_CHANNEL_MAX; i++) {
 438		struct netvsc_channel *nvchan = &net_device->chan_table[i];
 439
 440		nvchan->recv_buf = kzalloc(net_device->recv_section_size, GFP_KERNEL);
 441		if (nvchan->recv_buf == NULL) {
 442			ret = -ENOMEM;
 443			goto cleanup;
 444		}
 445	}
 446
 447	/* Setup receive completion ring.
 448	 * Add 1 to the recv_section_cnt because at least one entry in a
 449	 * ring buffer has to be empty.
 450	 */
 451	net_device->recv_completion_cnt = net_device->recv_section_cnt + 1;
 452	ret = netvsc_alloc_recv_comp_ring(net_device, 0);
 453	if (ret)
 454		goto cleanup;
 455
 456	/* Now setup the send buffer. */
 457	buf_size = device_info->send_sections * device_info->send_section_size;
 458	buf_size = round_up(buf_size, PAGE_SIZE);
 459
 460	net_device->send_buf = vzalloc(buf_size);
 461	if (!net_device->send_buf) {
 462		netdev_err(ndev, "unable to allocate send buffer of size %u\n",
 463			   buf_size);
 464		ret = -ENOMEM;
 465		goto cleanup;
 466	}
 467	net_device->send_buf_size = buf_size;
 468
 469	/* Establish the gpadl handle for this buffer on this
 470	 * channel.  Note: This call uses the vmbus connection rather
 471	 * than the channel to establish the gpadl handle.
 472	 */
 473	ret = vmbus_establish_gpadl(device->channel, net_device->send_buf,
 474				    buf_size,
 475				    &net_device->send_buf_gpadl_handle);
 476	if (ret != 0) {
 477		netdev_err(ndev,
 478			   "unable to establish send buffer's gpadl\n");
 479		goto cleanup;
 480	}
 481
 482	/* Notify the NetVsp of the gpadl handle */
 483	init_packet = &net_device->channel_init_pkt;
 484	memset(init_packet, 0, sizeof(struct nvsp_message));
 485	init_packet->hdr.msg_type = NVSP_MSG1_TYPE_SEND_SEND_BUF;
 486	init_packet->msg.v1_msg.send_send_buf.gpadl_handle =
 487		net_device->send_buf_gpadl_handle.gpadl_handle;
 488	init_packet->msg.v1_msg.send_send_buf.id = NETVSC_SEND_BUFFER_ID;
 489
 490	trace_nvsp_send(ndev, init_packet);
 491
 492	/* Send the gpadl notification request */
 493	ret = vmbus_sendpacket(device->channel, init_packet,
 494			       sizeof(struct nvsp_message),
 495			       (unsigned long)init_packet,
 496			       VM_PKT_DATA_INBAND,
 497			       VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
 498	if (ret != 0) {
 499		netdev_err(ndev,
 500			   "unable to send send buffer's gpadl to netvsp\n");
 501		goto cleanup;
 502	}
 503
 504	wait_for_completion(&net_device->channel_init_wait);
 505
 506	/* Check the response */
 507	if (init_packet->msg.v1_msg.
 508	    send_send_buf_complete.status != NVSP_STAT_SUCCESS) {
 509		netdev_err(ndev, "Unable to complete send buffer "
 510			   "initialization with NetVsp - status %d\n",
 511			   init_packet->msg.v1_msg.
 512			   send_send_buf_complete.status);
 513		ret = -EINVAL;
 514		goto cleanup;
 515	}
 516
 517	/* Parse the response */
 518	net_device->send_section_size = init_packet->msg.
 519				v1_msg.send_send_buf_complete.section_size;
 520	if (net_device->send_section_size < NETVSC_MTU_MIN) {
 521		netdev_err(ndev, "invalid send_section_size %u\n",
 522			   net_device->send_section_size);
 523		ret = -EINVAL;
 524		goto cleanup;
 525	}
 526
 527	/* Section count is simply the size divided by the section size. */
 528	net_device->send_section_cnt = buf_size / net_device->send_section_size;
 529
 530	netdev_dbg(ndev, "Send section size: %d, Section count:%d\n",
 531		   net_device->send_section_size, net_device->send_section_cnt);
 532
 533	/* Setup state for managing the send buffer. */
 534	net_device->send_section_map = bitmap_zalloc(net_device->send_section_cnt,
 535						     GFP_KERNEL);
 536	if (!net_device->send_section_map) {
 
 537		ret = -ENOMEM;
 538		goto cleanup;
 539	}
 540
 541	goto exit;
 542
 543cleanup:
 544	netvsc_revoke_recv_buf(device, net_device, ndev);
 545	netvsc_revoke_send_buf(device, net_device, ndev);
 546	netvsc_teardown_recv_gpadl(device, net_device, ndev);
 547	netvsc_teardown_send_gpadl(device, net_device, ndev);
 548
 549exit:
 550	return ret;
 551}
 552
 553/* Negotiate NVSP protocol version */
 554static int negotiate_nvsp_ver(struct hv_device *device,
 555			      struct netvsc_device *net_device,
 556			      struct nvsp_message *init_packet,
 557			      u32 nvsp_ver)
 558{
 559	struct net_device *ndev = hv_get_drvdata(device);
 560	int ret;
 561
 562	memset(init_packet, 0, sizeof(struct nvsp_message));
 563	init_packet->hdr.msg_type = NVSP_MSG_TYPE_INIT;
 564	init_packet->msg.init_msg.init.min_protocol_ver = nvsp_ver;
 565	init_packet->msg.init_msg.init.max_protocol_ver = nvsp_ver;
 566	trace_nvsp_send(ndev, init_packet);
 567
 568	/* Send the init request */
 569	ret = vmbus_sendpacket(device->channel, init_packet,
 570			       sizeof(struct nvsp_message),
 571			       (unsigned long)init_packet,
 572			       VM_PKT_DATA_INBAND,
 573			       VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
 574
 575	if (ret != 0)
 576		return ret;
 577
 578	wait_for_completion(&net_device->channel_init_wait);
 579
 580	if (init_packet->msg.init_msg.init_complete.status !=
 581	    NVSP_STAT_SUCCESS)
 582		return -EINVAL;
 583
 584	if (nvsp_ver == NVSP_PROTOCOL_VERSION_1)
 585		return 0;
 586
 587	/* NVSPv2 or later: Send NDIS config */
 588	memset(init_packet, 0, sizeof(struct nvsp_message));
 589	init_packet->hdr.msg_type = NVSP_MSG2_TYPE_SEND_NDIS_CONFIG;
 590	init_packet->msg.v2_msg.send_ndis_config.mtu = ndev->mtu + ETH_HLEN;
 591	init_packet->msg.v2_msg.send_ndis_config.capability.ieee8021q = 1;
 592
 593	if (nvsp_ver >= NVSP_PROTOCOL_VERSION_5) {
 594		if (hv_is_isolation_supported())
 595			netdev_info(ndev, "SR-IOV not advertised by guests on the host supporting isolation\n");
 596		else
 597			init_packet->msg.v2_msg.send_ndis_config.capability.sriov = 1;
 598
 599		/* Teaming bit is needed to receive link speed updates */
 600		init_packet->msg.v2_msg.send_ndis_config.capability.teaming = 1;
 601	}
 602
 603	if (nvsp_ver >= NVSP_PROTOCOL_VERSION_61)
 604		init_packet->msg.v2_msg.send_ndis_config.capability.rsc = 1;
 605
 606	trace_nvsp_send(ndev, init_packet);
 607
 608	ret = vmbus_sendpacket(device->channel, init_packet,
 609				sizeof(struct nvsp_message),
 610				VMBUS_RQST_ID_NO_RESPONSE,
 611				VM_PKT_DATA_INBAND, 0);
 612
 613	return ret;
 614}
 615
 616static int netvsc_connect_vsp(struct hv_device *device,
 617			      struct netvsc_device *net_device,
 618			      const struct netvsc_device_info *device_info)
 619{
 620	struct net_device *ndev = hv_get_drvdata(device);
 621	static const u32 ver_list[] = {
 622		NVSP_PROTOCOL_VERSION_1, NVSP_PROTOCOL_VERSION_2,
 623		NVSP_PROTOCOL_VERSION_4, NVSP_PROTOCOL_VERSION_5,
 624		NVSP_PROTOCOL_VERSION_6, NVSP_PROTOCOL_VERSION_61
 625	};
 626	struct nvsp_message *init_packet;
 627	int ndis_version, i, ret;
 628
 629	init_packet = &net_device->channel_init_pkt;
 630
 631	/* Negotiate the latest NVSP protocol supported */
 632	for (i = ARRAY_SIZE(ver_list) - 1; i >= 0; i--)
 633		if (negotiate_nvsp_ver(device, net_device, init_packet,
 634				       ver_list[i])  == 0) {
 635			net_device->nvsp_version = ver_list[i];
 636			break;
 637		}
 638
 639	if (i < 0) {
 640		ret = -EPROTO;
 641		goto cleanup;
 642	}
 643
 644	if (hv_is_isolation_supported() && net_device->nvsp_version < NVSP_PROTOCOL_VERSION_61) {
 645		netdev_err(ndev, "Invalid NVSP version 0x%x (expected >= 0x%x) from the host supporting isolation\n",
 646			   net_device->nvsp_version, NVSP_PROTOCOL_VERSION_61);
 647		ret = -EPROTO;
 648		goto cleanup;
 649	}
 650
 651	pr_debug("Negotiated NVSP version:%x\n", net_device->nvsp_version);
 652
 653	/* Send the ndis version */
 654	memset(init_packet, 0, sizeof(struct nvsp_message));
 655
 656	if (net_device->nvsp_version <= NVSP_PROTOCOL_VERSION_4)
 657		ndis_version = 0x00060001;
 658	else
 659		ndis_version = 0x0006001e;
 660
 661	init_packet->hdr.msg_type = NVSP_MSG1_TYPE_SEND_NDIS_VER;
 662	init_packet->msg.v1_msg.
 663		send_ndis_ver.ndis_major_ver =
 664				(ndis_version & 0xFFFF0000) >> 16;
 665	init_packet->msg.v1_msg.
 666		send_ndis_ver.ndis_minor_ver =
 667				ndis_version & 0xFFFF;
 668
 669	trace_nvsp_send(ndev, init_packet);
 670
 671	/* Send the init request */
 672	ret = vmbus_sendpacket(device->channel, init_packet,
 673				sizeof(struct nvsp_message),
 674				VMBUS_RQST_ID_NO_RESPONSE,
 675				VM_PKT_DATA_INBAND, 0);
 676	if (ret != 0)
 677		goto cleanup;
 678
 679
 680	ret = netvsc_init_buf(device, net_device, device_info);
 681
 682cleanup:
 683	return ret;
 684}
 685
 686/*
 687 * netvsc_device_remove - Callback when the root bus device is removed
 688 */
 689void netvsc_device_remove(struct hv_device *device)
 690{
 691	struct net_device *ndev = hv_get_drvdata(device);
 692	struct net_device_context *net_device_ctx = netdev_priv(ndev);
 693	struct netvsc_device *net_device
 694		= rtnl_dereference(net_device_ctx->nvdev);
 695	int i;
 696
 697	/*
 698	 * Revoke receive buffer. If host is pre-Win2016 then tear down
 699	 * receive buffer GPADL. Do the same for send buffer.
 700	 */
 701	netvsc_revoke_recv_buf(device, net_device, ndev);
 702	if (vmbus_proto_version < VERSION_WIN10)
 703		netvsc_teardown_recv_gpadl(device, net_device, ndev);
 704
 705	netvsc_revoke_send_buf(device, net_device, ndev);
 706	if (vmbus_proto_version < VERSION_WIN10)
 707		netvsc_teardown_send_gpadl(device, net_device, ndev);
 708
 709	RCU_INIT_POINTER(net_device_ctx->nvdev, NULL);
 710
 711	/* Disable NAPI and disassociate its context from the device. */
 712	for (i = 0; i < net_device->num_chn; i++) {
 713		/* See also vmbus_reset_channel_cb(). */
 714		/* only disable enabled NAPI channel */
 715		if (i < ndev->real_num_rx_queues) {
 716			netif_queue_set_napi(ndev, i, NETDEV_QUEUE_TYPE_TX,
 717					     NULL);
 718			netif_queue_set_napi(ndev, i, NETDEV_QUEUE_TYPE_RX,
 719					     NULL);
 720			napi_disable(&net_device->chan_table[i].napi);
 721		}
 722
 723		netif_napi_del(&net_device->chan_table[i].napi);
 724	}
 725
 726	/*
 727	 * At this point, no one should be accessing net_device
 728	 * except in here
 729	 */
 730	netdev_dbg(ndev, "net device safe to remove\n");
 731
 732	/* Now, we can close the channel safely */
 733	vmbus_close(device->channel);
 734
 735	/*
 736	 * If host is Win2016 or higher then we do the GPADL tear down
 737	 * here after VMBus is closed.
 738	*/
 739	if (vmbus_proto_version >= VERSION_WIN10) {
 740		netvsc_teardown_recv_gpadl(device, net_device, ndev);
 741		netvsc_teardown_send_gpadl(device, net_device, ndev);
 742	}
 743
 744	/* Release all resources */
 745	free_netvsc_device_rcu(net_device);
 746}
 747
 748#define RING_AVAIL_PERCENT_HIWATER 20
 749#define RING_AVAIL_PERCENT_LOWATER 10
 750
 
 
 
 
 
 
 
 
 
 
 
 751static inline void netvsc_free_send_slot(struct netvsc_device *net_device,
 752					 u32 index)
 753{
 754	sync_change_bit(index, net_device->send_section_map);
 755}
 756
 757static void netvsc_send_tx_complete(struct net_device *ndev,
 758				    struct netvsc_device *net_device,
 759				    struct vmbus_channel *channel,
 760				    const struct vmpacket_descriptor *desc,
 761				    int budget)
 762{
 
 
 763	struct net_device_context *ndev_ctx = netdev_priv(ndev);
 764	struct sk_buff *skb;
 765	u16 q_idx = 0;
 766	int queue_sends;
 767	u64 cmd_rqst;
 768
 769	cmd_rqst = channel->request_addr_callback(channel, desc->trans_id);
 770	if (cmd_rqst == VMBUS_RQST_ERROR) {
 771		netdev_err(ndev, "Invalid transaction ID %llx\n", desc->trans_id);
 772		return;
 773	}
 774
 775	skb = (struct sk_buff *)(unsigned long)cmd_rqst;
 776
 777	/* Notify the layer above us */
 778	if (likely(skb)) {
 779		struct hv_netvsc_packet *packet
 780			= (struct hv_netvsc_packet *)skb->cb;
 781		u32 send_index = packet->send_buf_index;
 782		struct netvsc_stats_tx *tx_stats;
 783
 784		if (send_index != NETVSC_INVALID_INDEX)
 785			netvsc_free_send_slot(net_device, send_index);
 786		q_idx = packet->q_idx;
 
 787
 788		tx_stats = &net_device->chan_table[q_idx].tx_stats;
 789
 790		u64_stats_update_begin(&tx_stats->syncp);
 791		tx_stats->packets += packet->total_packets;
 792		tx_stats->bytes += packet->total_bytes;
 793		u64_stats_update_end(&tx_stats->syncp);
 794
 795		netvsc_dma_unmap(ndev_ctx->device_ctx, packet);
 796		napi_consume_skb(skb, budget);
 797	}
 798
 799	queue_sends =
 800		atomic_dec_return(&net_device->chan_table[q_idx].queue_sends);
 801
 802	if (unlikely(net_device->destroy)) {
 803		if (queue_sends == 0)
 804			wake_up(&net_device->wait_drain);
 805	} else {
 806		struct netdev_queue *txq = netdev_get_tx_queue(ndev, q_idx);
 807
 808		if (netif_tx_queue_stopped(txq) && !net_device->tx_disable &&
 809		    (hv_get_avail_to_write_percent(&channel->outbound) >
 810		     RING_AVAIL_PERCENT_HIWATER || queue_sends < 1)) {
 811			netif_tx_wake_queue(txq);
 812			ndev_ctx->eth_stats.wake_queue++;
 813		}
 814	}
 815}
 816
 817static void netvsc_send_completion(struct net_device *ndev,
 818				   struct netvsc_device *net_device,
 819				   struct vmbus_channel *incoming_channel,
 
 820				   const struct vmpacket_descriptor *desc,
 821				   int budget)
 822{
 823	const struct nvsp_message *nvsp_packet;
 824	u32 msglen = hv_pkt_datalen(desc);
 825	struct nvsp_message *pkt_rqst;
 826	u64 cmd_rqst;
 827	u32 status;
 828
 829	/* First check if this is a VMBUS completion without data payload */
 830	if (!msglen) {
 831		cmd_rqst = incoming_channel->request_addr_callback(incoming_channel,
 832								   desc->trans_id);
 833		if (cmd_rqst == VMBUS_RQST_ERROR) {
 834			netdev_err(ndev, "Invalid transaction ID %llx\n", desc->trans_id);
 835			return;
 836		}
 837
 838		pkt_rqst = (struct nvsp_message *)(uintptr_t)cmd_rqst;
 839		switch (pkt_rqst->hdr.msg_type) {
 840		case NVSP_MSG4_TYPE_SWITCH_DATA_PATH:
 841			complete(&net_device->channel_init_wait);
 842			break;
 843
 844		default:
 845			netdev_err(ndev, "Unexpected VMBUS completion!!\n");
 846		}
 847		return;
 848	}
 849
 850	/* Ensure packet is big enough to read header fields */
 851	if (msglen < sizeof(struct nvsp_message_header)) {
 852		netdev_err(ndev, "nvsp_message length too small: %u\n", msglen);
 853		return;
 854	}
 855
 856	nvsp_packet = hv_pkt_data(desc);
 857	switch (nvsp_packet->hdr.msg_type) {
 858	case NVSP_MSG_TYPE_INIT_COMPLETE:
 859		if (msglen < sizeof(struct nvsp_message_header) +
 860				sizeof(struct nvsp_message_init_complete)) {
 861			netdev_err(ndev, "nvsp_msg length too small: %u\n",
 862				   msglen);
 863			return;
 864		}
 865		break;
 866
 867	case NVSP_MSG1_TYPE_SEND_RECV_BUF_COMPLETE:
 868		if (msglen < sizeof(struct nvsp_message_header) +
 869				sizeof(struct nvsp_1_message_send_receive_buffer_complete)) {
 870			netdev_err(ndev, "nvsp_msg1 length too small: %u\n",
 871				   msglen);
 872			return;
 873		}
 874		break;
 875
 876	case NVSP_MSG1_TYPE_SEND_SEND_BUF_COMPLETE:
 877		if (msglen < sizeof(struct nvsp_message_header) +
 878				sizeof(struct nvsp_1_message_send_send_buffer_complete)) {
 879			netdev_err(ndev, "nvsp_msg1 length too small: %u\n",
 880				   msglen);
 881			return;
 882		}
 883		break;
 884
 885	case NVSP_MSG5_TYPE_SUBCHANNEL:
 886		if (msglen < sizeof(struct nvsp_message_header) +
 887				sizeof(struct nvsp_5_subchannel_complete)) {
 888			netdev_err(ndev, "nvsp_msg5 length too small: %u\n",
 889				   msglen);
 890			return;
 891		}
 892		break;
 893
 894	case NVSP_MSG1_TYPE_SEND_RNDIS_PKT_COMPLETE:
 895		if (msglen < sizeof(struct nvsp_message_header) +
 896		    sizeof(struct nvsp_1_message_send_rndis_packet_complete)) {
 897			if (net_ratelimit())
 898				netdev_err(ndev, "nvsp_rndis_pkt_complete length too small: %u\n",
 899					   msglen);
 900			return;
 901		}
 902
 903		/* If status indicates an error, output a message so we know
 904		 * there's a problem. But process the completion anyway so the
 905		 * resources are released.
 906		 */
 907		status = nvsp_packet->msg.v1_msg.send_rndis_pkt_complete.status;
 908		if (status != NVSP_STAT_SUCCESS && net_ratelimit())
 909			netdev_err(ndev, "nvsp_rndis_pkt_complete error status: %x\n",
 910				   status);
 911
 912		netvsc_send_tx_complete(ndev, net_device, incoming_channel,
 913					desc, budget);
 914		return;
 915
 916	default:
 917		netdev_err(ndev,
 918			   "Unknown send completion type %d received!!\n",
 919			   nvsp_packet->hdr.msg_type);
 920		return;
 921	}
 922
 923	/* Copy the response back */
 924	memcpy(&net_device->channel_init_pkt, nvsp_packet,
 925	       sizeof(struct nvsp_message));
 926	complete(&net_device->channel_init_wait);
 927}
 928
 929static u32 netvsc_get_next_send_section(struct netvsc_device *net_device)
 930{
 931	unsigned long *map_addr = net_device->send_section_map;
 932	unsigned int i;
 933
 934	for_each_clear_bit(i, map_addr, net_device->send_section_cnt) {
 935		if (sync_test_and_set_bit(i, map_addr) == 0)
 936			return i;
 937	}
 938
 939	return NETVSC_INVALID_INDEX;
 940}
 941
 942static void netvsc_copy_to_send_buf(struct netvsc_device *net_device,
 943				    unsigned int section_index,
 944				    u32 pend_size,
 945				    struct hv_netvsc_packet *packet,
 946				    struct rndis_message *rndis_msg,
 947				    struct hv_page_buffer *pb,
 948				    bool xmit_more)
 949{
 950	char *start = net_device->send_buf;
 951	char *dest = start + (section_index * net_device->send_section_size)
 952		     + pend_size;
 953	int i;
 954	u32 padding = 0;
 955	u32 page_count = packet->cp_partial ? packet->rmsg_pgcnt :
 956		packet->page_buf_cnt;
 957	u32 remain;
 958
 959	/* Add padding */
 960	remain = packet->total_data_buflen & (net_device->pkt_align - 1);
 961	if (xmit_more && remain) {
 962		padding = net_device->pkt_align - remain;
 963		rndis_msg->msg_len += padding;
 964		packet->total_data_buflen += padding;
 965	}
 966
 967	for (i = 0; i < page_count; i++) {
 968		char *src = phys_to_virt(pb[i].pfn << HV_HYP_PAGE_SHIFT);
 969		u32 offset = pb[i].offset;
 970		u32 len = pb[i].len;
 971
 972		memcpy(dest, (src + offset), len);
 973		dest += len;
 974	}
 975
 976	if (padding)
 977		memset(dest, 0, padding);
 978}
 979
 980void netvsc_dma_unmap(struct hv_device *hv_dev,
 981		      struct hv_netvsc_packet *packet)
 982{
 983	int i;
 984
 985	if (!hv_is_isolation_supported())
 986		return;
 987
 988	if (!packet->dma_range)
 989		return;
 990
 991	for (i = 0; i < packet->page_buf_cnt; i++)
 992		dma_unmap_single(&hv_dev->device, packet->dma_range[i].dma,
 993				 packet->dma_range[i].mapping_size,
 994				 DMA_TO_DEVICE);
 995
 996	kfree(packet->dma_range);
 997}
 998
 999/* netvsc_dma_map - Map swiotlb bounce buffer with data page of
1000 * packet sent by vmbus_sendpacket_pagebuffer() in the Isolation
1001 * VM.
1002 *
1003 * In isolation VM, netvsc send buffer has been marked visible to
1004 * host and so the data copied to send buffer doesn't need to use
1005 * bounce buffer. The data pages handled by vmbus_sendpacket_pagebuffer()
1006 * may not be copied to send buffer and so these pages need to be
1007 * mapped with swiotlb bounce buffer. netvsc_dma_map() is to do
1008 * that. The pfns in the struct hv_page_buffer need to be converted
1009 * to bounce buffer's pfn. The loop here is necessary because the
1010 * entries in the page buffer array are not necessarily full
1011 * pages of data.  Each entry in the array has a separate offset and
1012 * len that may be non-zero, even for entries in the middle of the
1013 * array.  And the entries are not physically contiguous.  So each
1014 * entry must be individually mapped rather than as a contiguous unit.
1015 * So not use dma_map_sg() here.
1016 */
1017static int netvsc_dma_map(struct hv_device *hv_dev,
1018			  struct hv_netvsc_packet *packet,
1019			  struct hv_page_buffer *pb)
1020{
1021	u32 page_count = packet->page_buf_cnt;
1022	dma_addr_t dma;
1023	int i;
1024
1025	if (!hv_is_isolation_supported())
1026		return 0;
1027
1028	packet->dma_range = kcalloc(page_count,
1029				    sizeof(*packet->dma_range),
1030				    GFP_ATOMIC);
1031	if (!packet->dma_range)
1032		return -ENOMEM;
1033
1034	for (i = 0; i < page_count; i++) {
1035		char *src = phys_to_virt((pb[i].pfn << HV_HYP_PAGE_SHIFT)
1036					 + pb[i].offset);
1037		u32 len = pb[i].len;
1038
1039		dma = dma_map_single(&hv_dev->device, src, len,
1040				     DMA_TO_DEVICE);
1041		if (dma_mapping_error(&hv_dev->device, dma)) {
1042			kfree(packet->dma_range);
1043			return -ENOMEM;
1044		}
1045
1046		/* pb[].offset and pb[].len are not changed during dma mapping
1047		 * and so not reassign.
1048		 */
1049		packet->dma_range[i].dma = dma;
1050		packet->dma_range[i].mapping_size = len;
1051		pb[i].pfn = dma >> HV_HYP_PAGE_SHIFT;
1052	}
1053
1054	return 0;
1055}
1056
1057static inline int netvsc_send_pkt(
1058	struct hv_device *device,
1059	struct hv_netvsc_packet *packet,
1060	struct netvsc_device *net_device,
1061	struct hv_page_buffer *pb,
1062	struct sk_buff *skb)
1063{
1064	struct nvsp_message nvmsg;
1065	struct nvsp_1_message_send_rndis_packet *rpkt =
1066		&nvmsg.msg.v1_msg.send_rndis_pkt;
1067	struct netvsc_channel * const nvchan =
1068		&net_device->chan_table[packet->q_idx];
1069	struct vmbus_channel *out_channel = nvchan->channel;
1070	struct net_device *ndev = hv_get_drvdata(device);
1071	struct net_device_context *ndev_ctx = netdev_priv(ndev);
1072	struct netdev_queue *txq = netdev_get_tx_queue(ndev, packet->q_idx);
1073	u64 req_id;
1074	int ret;
1075	u32 ring_avail = hv_get_avail_to_write_percent(&out_channel->outbound);
1076
1077	memset(&nvmsg, 0, sizeof(struct nvsp_message));
1078	nvmsg.hdr.msg_type = NVSP_MSG1_TYPE_SEND_RNDIS_PKT;
1079	if (skb)
1080		rpkt->channel_type = 0;		/* 0 is RMC_DATA */
1081	else
1082		rpkt->channel_type = 1;		/* 1 is RMC_CONTROL */
1083
1084	rpkt->send_buf_section_index = packet->send_buf_index;
1085	if (packet->send_buf_index == NETVSC_INVALID_INDEX)
1086		rpkt->send_buf_section_size = 0;
1087	else
1088		rpkt->send_buf_section_size = packet->total_data_buflen;
1089
1090	req_id = (ulong)skb;
1091
1092	if (out_channel->rescind)
1093		return -ENODEV;
1094
1095	trace_nvsp_send_pkt(ndev, out_channel, rpkt);
1096
1097	packet->dma_range = NULL;
1098	if (packet->page_buf_cnt) {
1099		if (packet->cp_partial)
1100			pb += packet->rmsg_pgcnt;
1101
1102		ret = netvsc_dma_map(ndev_ctx->device_ctx, packet, pb);
1103		if (ret) {
1104			ret = -EAGAIN;
1105			goto exit;
1106		}
1107
1108		ret = vmbus_sendpacket_pagebuffer(out_channel,
1109						  pb, packet->page_buf_cnt,
1110						  &nvmsg, sizeof(nvmsg),
1111						  req_id);
1112
1113		if (ret)
1114			netvsc_dma_unmap(ndev_ctx->device_ctx, packet);
1115	} else {
1116		ret = vmbus_sendpacket(out_channel,
1117				       &nvmsg, sizeof(nvmsg),
1118				       req_id, VM_PKT_DATA_INBAND,
1119				       VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
1120	}
1121
1122exit:
1123	if (ret == 0) {
1124		atomic_inc_return(&nvchan->queue_sends);
1125
1126		if (ring_avail < RING_AVAIL_PERCENT_LOWATER) {
1127			netif_tx_stop_queue(txq);
1128			ndev_ctx->eth_stats.stop_queue++;
1129		}
1130	} else if (ret == -EAGAIN) {
1131		netif_tx_stop_queue(txq);
1132		ndev_ctx->eth_stats.stop_queue++;
 
 
 
 
 
1133	} else {
1134		netdev_err(ndev,
1135			   "Unable to send packet pages %u len %u, ret %d\n",
1136			   packet->page_buf_cnt, packet->total_data_buflen,
1137			   ret);
1138	}
1139
1140	if (netif_tx_queue_stopped(txq) &&
1141	    atomic_read(&nvchan->queue_sends) < 1 &&
1142	    !net_device->tx_disable) {
1143		netif_tx_wake_queue(txq);
1144		ndev_ctx->eth_stats.wake_queue++;
1145		if (ret == -EAGAIN)
1146			ret = -ENOSPC;
1147	}
1148
1149	return ret;
1150}
1151
1152/* Move packet out of multi send data (msd), and clear msd */
1153static inline void move_pkt_msd(struct hv_netvsc_packet **msd_send,
1154				struct sk_buff **msd_skb,
1155				struct multi_send_data *msdp)
1156{
1157	*msd_skb = msdp->skb;
1158	*msd_send = msdp->pkt;
1159	msdp->skb = NULL;
1160	msdp->pkt = NULL;
1161	msdp->count = 0;
1162}
1163
1164/* RCU already held by caller */
1165/* Batching/bouncing logic is designed to attempt to optimize
1166 * performance.
1167 *
1168 * For small, non-LSO packets we copy the packet to a send buffer
1169 * which is pre-registered with the Hyper-V side. This enables the
1170 * hypervisor to avoid remapping the aperture to access the packet
1171 * descriptor and data.
1172 *
1173 * If we already started using a buffer and the netdev is transmitting
1174 * a burst of packets, keep on copying into the buffer until it is
1175 * full or we are done collecting a burst. If there is an existing
1176 * buffer with space for the RNDIS descriptor but not the packet, copy
1177 * the RNDIS descriptor to the buffer, keeping the packet in place.
1178 *
1179 * If we do batching and send more than one packet using a single
1180 * NetVSC message, free the SKBs of the packets copied, except for the
1181 * last packet. This is done to streamline the handling of the case
1182 * where the last packet only had the RNDIS descriptor copied to the
1183 * send buffer, with the data pointers included in the NetVSC message.
1184 */
1185int netvsc_send(struct net_device *ndev,
1186		struct hv_netvsc_packet *packet,
1187		struct rndis_message *rndis_msg,
1188		struct hv_page_buffer *pb,
1189		struct sk_buff *skb,
1190		bool xdp_tx)
1191{
1192	struct net_device_context *ndev_ctx = netdev_priv(ndev);
1193	struct netvsc_device *net_device
1194		= rcu_dereference_bh(ndev_ctx->nvdev);
1195	struct hv_device *device = ndev_ctx->device_ctx;
1196	int ret = 0;
1197	struct netvsc_channel *nvchan;
1198	u32 pktlen = packet->total_data_buflen, msd_len = 0;
1199	unsigned int section_index = NETVSC_INVALID_INDEX;
1200	struct multi_send_data *msdp;
1201	struct hv_netvsc_packet *msd_send = NULL, *cur_send = NULL;
1202	struct sk_buff *msd_skb = NULL;
1203	bool try_batch, xmit_more;
1204
1205	/* If device is rescinded, return error and packet will get dropped. */
1206	if (unlikely(!net_device || net_device->destroy))
1207		return -ENODEV;
1208
1209	nvchan = &net_device->chan_table[packet->q_idx];
1210	packet->send_buf_index = NETVSC_INVALID_INDEX;
1211	packet->cp_partial = false;
1212
1213	/* Send a control message or XDP packet directly without accessing
1214	 * msd (Multi-Send Data) field which may be changed during data packet
1215	 * processing.
1216	 */
1217	if (!skb || xdp_tx)
1218		return netvsc_send_pkt(device, packet, net_device, pb, skb);
1219
1220	/* batch packets in send buffer if possible */
1221	msdp = &nvchan->msd;
1222	if (msdp->pkt)
1223		msd_len = msdp->pkt->total_data_buflen;
1224
1225	try_batch =  msd_len > 0 && msdp->count < net_device->max_pkt;
1226	if (try_batch && msd_len + pktlen + net_device->pkt_align <
1227	    net_device->send_section_size) {
1228		section_index = msdp->pkt->send_buf_index;
1229
1230	} else if (try_batch && msd_len + packet->rmsg_size <
1231		   net_device->send_section_size) {
1232		section_index = msdp->pkt->send_buf_index;
1233		packet->cp_partial = true;
1234
1235	} else if (pktlen + net_device->pkt_align <
1236		   net_device->send_section_size) {
1237		section_index = netvsc_get_next_send_section(net_device);
1238		if (unlikely(section_index == NETVSC_INVALID_INDEX)) {
1239			++ndev_ctx->eth_stats.tx_send_full;
1240		} else {
1241			move_pkt_msd(&msd_send, &msd_skb, msdp);
1242			msd_len = 0;
1243		}
1244	}
1245
1246	/* Keep aggregating only if stack says more data is coming
1247	 * and not doing mixed modes send and not flow blocked
1248	 */
1249	xmit_more = netdev_xmit_more() &&
1250		!packet->cp_partial &&
1251		!netif_xmit_stopped(netdev_get_tx_queue(ndev, packet->q_idx));
1252
1253	if (section_index != NETVSC_INVALID_INDEX) {
1254		netvsc_copy_to_send_buf(net_device,
1255					section_index, msd_len,
1256					packet, rndis_msg, pb, xmit_more);
1257
1258		packet->send_buf_index = section_index;
1259
1260		if (packet->cp_partial) {
1261			packet->page_buf_cnt -= packet->rmsg_pgcnt;
1262			packet->total_data_buflen = msd_len + packet->rmsg_size;
1263		} else {
1264			packet->page_buf_cnt = 0;
1265			packet->total_data_buflen += msd_len;
1266		}
1267
1268		if (msdp->pkt) {
1269			packet->total_packets += msdp->pkt->total_packets;
1270			packet->total_bytes += msdp->pkt->total_bytes;
1271		}
1272
1273		if (msdp->skb)
1274			dev_consume_skb_any(msdp->skb);
1275
1276		if (xmit_more) {
1277			msdp->skb = skb;
1278			msdp->pkt = packet;
1279			msdp->count++;
1280		} else {
1281			cur_send = packet;
1282			msdp->skb = NULL;
1283			msdp->pkt = NULL;
1284			msdp->count = 0;
1285		}
1286	} else {
1287		move_pkt_msd(&msd_send, &msd_skb, msdp);
1288		cur_send = packet;
1289	}
1290
1291	if (msd_send) {
1292		int m_ret = netvsc_send_pkt(device, msd_send, net_device,
1293					    NULL, msd_skb);
1294
1295		if (m_ret != 0) {
1296			netvsc_free_send_slot(net_device,
1297					      msd_send->send_buf_index);
1298			dev_kfree_skb_any(msd_skb);
1299		}
1300	}
1301
1302	if (cur_send)
1303		ret = netvsc_send_pkt(device, cur_send, net_device, pb, skb);
1304
1305	if (ret != 0 && section_index != NETVSC_INVALID_INDEX)
1306		netvsc_free_send_slot(net_device, section_index);
1307
1308	return ret;
1309}
1310
1311/* Send pending recv completions */
1312static int send_recv_completions(struct net_device *ndev,
1313				 struct netvsc_device *nvdev,
1314				 struct netvsc_channel *nvchan)
1315{
1316	struct multi_recv_comp *mrc = &nvchan->mrc;
1317	struct recv_comp_msg {
1318		struct nvsp_message_header hdr;
1319		u32 status;
1320	}  __packed;
1321	struct recv_comp_msg msg = {
1322		.hdr.msg_type = NVSP_MSG1_TYPE_SEND_RNDIS_PKT_COMPLETE,
1323	};
1324	int ret;
1325
1326	while (mrc->first != mrc->next) {
1327		const struct recv_comp_data *rcd
1328			= mrc->slots + mrc->first;
1329
1330		msg.status = rcd->status;
1331		ret = vmbus_sendpacket(nvchan->channel, &msg, sizeof(msg),
1332				       rcd->tid, VM_PKT_COMP, 0);
1333		if (unlikely(ret)) {
1334			struct net_device_context *ndev_ctx = netdev_priv(ndev);
1335
1336			++ndev_ctx->eth_stats.rx_comp_busy;
1337			return ret;
1338		}
1339
1340		if (++mrc->first == nvdev->recv_completion_cnt)
1341			mrc->first = 0;
1342	}
1343
1344	/* receive completion ring has been emptied */
1345	if (unlikely(nvdev->destroy))
1346		wake_up(&nvdev->wait_drain);
1347
1348	return 0;
1349}
1350
1351/* Count how many receive completions are outstanding */
1352static void recv_comp_slot_avail(const struct netvsc_device *nvdev,
1353				 const struct multi_recv_comp *mrc,
1354				 u32 *filled, u32 *avail)
1355{
1356	u32 count = nvdev->recv_completion_cnt;
1357
1358	if (mrc->next >= mrc->first)
1359		*filled = mrc->next - mrc->first;
1360	else
1361		*filled = (count - mrc->first) + mrc->next;
1362
1363	*avail = count - *filled - 1;
1364}
1365
1366/* Add receive complete to ring to send to host. */
1367static void enq_receive_complete(struct net_device *ndev,
1368				 struct netvsc_device *nvdev, u16 q_idx,
1369				 u64 tid, u32 status)
1370{
1371	struct netvsc_channel *nvchan = &nvdev->chan_table[q_idx];
1372	struct multi_recv_comp *mrc = &nvchan->mrc;
1373	struct recv_comp_data *rcd;
1374	u32 filled, avail;
1375
1376	recv_comp_slot_avail(nvdev, mrc, &filled, &avail);
1377
1378	if (unlikely(filled > NAPI_POLL_WEIGHT)) {
1379		send_recv_completions(ndev, nvdev, nvchan);
1380		recv_comp_slot_avail(nvdev, mrc, &filled, &avail);
1381	}
1382
1383	if (unlikely(!avail)) {
1384		netdev_err(ndev, "Recv_comp full buf q:%hd, tid:%llx\n",
1385			   q_idx, tid);
1386		return;
1387	}
1388
1389	rcd = mrc->slots + mrc->next;
1390	rcd->tid = tid;
1391	rcd->status = status;
1392
1393	if (++mrc->next == nvdev->recv_completion_cnt)
1394		mrc->next = 0;
1395}
1396
1397static int netvsc_receive(struct net_device *ndev,
1398			  struct netvsc_device *net_device,
1399			  struct netvsc_channel *nvchan,
1400			  const struct vmpacket_descriptor *desc)
 
 
 
1401{
1402	struct net_device_context *net_device_ctx = netdev_priv(ndev);
1403	struct vmbus_channel *channel = nvchan->channel;
1404	const struct vmtransfer_page_packet_header *vmxferpage_packet
1405		= container_of(desc, const struct vmtransfer_page_packet_header, d);
1406	const struct nvsp_message *nvsp = hv_pkt_data(desc);
1407	u32 msglen = hv_pkt_datalen(desc);
1408	u16 q_idx = channel->offermsg.offer.sub_channel_index;
1409	char *recv_buf = net_device->recv_buf;
1410	u32 status = NVSP_STAT_SUCCESS;
1411	int i;
1412	int count = 0;
1413
1414	/* Ensure packet is big enough to read header fields */
1415	if (msglen < sizeof(struct nvsp_message_header)) {
1416		netif_err(net_device_ctx, rx_err, ndev,
1417			  "invalid nvsp header, length too small: %u\n",
1418			  msglen);
1419		return 0;
1420	}
1421
1422	/* Make sure this is a valid nvsp packet */
1423	if (unlikely(nvsp->hdr.msg_type != NVSP_MSG1_TYPE_SEND_RNDIS_PKT)) {
1424		netif_err(net_device_ctx, rx_err, ndev,
1425			  "Unknown nvsp packet type received %u\n",
1426			  nvsp->hdr.msg_type);
1427		return 0;
1428	}
1429
1430	/* Validate xfer page pkt header */
1431	if ((desc->offset8 << 3) < sizeof(struct vmtransfer_page_packet_header)) {
1432		netif_err(net_device_ctx, rx_err, ndev,
1433			  "Invalid xfer page pkt, offset too small: %u\n",
1434			  desc->offset8 << 3);
1435		return 0;
1436	}
1437
1438	if (unlikely(vmxferpage_packet->xfer_pageset_id != NETVSC_RECEIVE_BUFFER_ID)) {
1439		netif_err(net_device_ctx, rx_err, ndev,
1440			  "Invalid xfer page set id - expecting %x got %x\n",
1441			  NETVSC_RECEIVE_BUFFER_ID,
1442			  vmxferpage_packet->xfer_pageset_id);
1443		return 0;
1444	}
1445
1446	count = vmxferpage_packet->range_cnt;
1447
1448	/* Check count for a valid value */
1449	if (NETVSC_XFER_HEADER_SIZE(count) > desc->offset8 << 3) {
1450		netif_err(net_device_ctx, rx_err, ndev,
1451			  "Range count is not valid: %d\n",
1452			  count);
1453		return 0;
1454	}
1455
1456	/* Each range represents 1 RNDIS pkt that contains 1 ethernet frame */
1457	for (i = 0; i < count; i++) {
1458		u32 offset = vmxferpage_packet->ranges[i].byte_offset;
1459		u32 buflen = vmxferpage_packet->ranges[i].byte_count;
1460		void *data;
1461		int ret;
1462
1463		if (unlikely(offset > net_device->recv_buf_size ||
1464			     buflen > net_device->recv_buf_size - offset)) {
1465			nvchan->rsc.cnt = 0;
1466			status = NVSP_STAT_FAIL;
1467			netif_err(net_device_ctx, rx_err, ndev,
1468				  "Packet offset:%u + len:%u too big\n",
1469				  offset, buflen);
1470
1471			continue;
1472		}
1473
1474		/* We're going to copy (sections of) the packet into nvchan->recv_buf;
1475		 * make sure that nvchan->recv_buf is large enough to hold the packet.
1476		 */
1477		if (unlikely(buflen > net_device->recv_section_size)) {
1478			nvchan->rsc.cnt = 0;
1479			status = NVSP_STAT_FAIL;
1480			netif_err(net_device_ctx, rx_err, ndev,
1481				  "Packet too big: buflen=%u recv_section_size=%u\n",
1482				  buflen, net_device->recv_section_size);
1483
1484			continue;
1485		}
1486
1487		data = recv_buf + offset;
1488
1489		nvchan->rsc.is_last = (i == count - 1);
1490
1491		trace_rndis_recv(ndev, q_idx, data);
1492
1493		/* Pass it to the upper layer */
1494		ret = rndis_filter_receive(ndev, net_device,
1495					   nvchan, data, buflen);
1496
1497		if (unlikely(ret != NVSP_STAT_SUCCESS)) {
1498			/* Drop incomplete packet */
1499			nvchan->rsc.cnt = 0;
1500			status = NVSP_STAT_FAIL;
1501		}
1502	}
1503
1504	enq_receive_complete(ndev, net_device, q_idx,
1505			     vmxferpage_packet->d.trans_id, status);
1506
1507	return count;
1508}
1509
1510static void netvsc_send_table(struct net_device *ndev,
1511			      struct netvsc_device *nvscdev,
1512			      const struct nvsp_message *nvmsg,
1513			      u32 msglen)
1514{
 
1515	struct net_device_context *net_device_ctx = netdev_priv(ndev);
1516	u32 count, offset, *tab;
1517	int i;
1518
1519	/* Ensure packet is big enough to read send_table fields */
1520	if (msglen < sizeof(struct nvsp_message_header) +
1521		     sizeof(struct nvsp_5_send_indirect_table)) {
1522		netdev_err(ndev, "nvsp_v5_msg length too small: %u\n", msglen);
1523		return;
1524	}
1525
1526	count = nvmsg->msg.v5_msg.send_table.count;
1527	offset = nvmsg->msg.v5_msg.send_table.offset;
1528
1529	if (count != VRSS_SEND_TAB_SIZE) {
1530		netdev_err(ndev, "Received wrong send-table size:%u\n", count);
1531		return;
1532	}
1533
1534	/* If negotiated version <= NVSP_PROTOCOL_VERSION_6, the offset may be
1535	 * wrong due to a host bug. So fix the offset here.
1536	 */
1537	if (nvscdev->nvsp_version <= NVSP_PROTOCOL_VERSION_6 &&
1538	    msglen >= sizeof(struct nvsp_message_header) +
1539	    sizeof(union nvsp_6_message_uber) + count * sizeof(u32))
1540		offset = sizeof(struct nvsp_message_header) +
1541			 sizeof(union nvsp_6_message_uber);
1542
1543	/* Boundary check for all versions */
1544	if (msglen < count * sizeof(u32) || offset > msglen - count * sizeof(u32)) {
1545		netdev_err(ndev, "Received send-table offset too big:%u\n",
1546			   offset);
1547		return;
1548	}
1549
1550	tab = (void *)nvmsg + offset;
1551
1552	for (i = 0; i < count; i++)
1553		net_device_ctx->tx_table[i] = tab[i];
1554}
1555
1556static void netvsc_send_vf(struct net_device *ndev,
1557			   const struct nvsp_message *nvmsg,
1558			   u32 msglen)
1559{
1560	struct net_device_context *net_device_ctx = netdev_priv(ndev);
1561
1562	/* Ensure packet is big enough to read its fields */
1563	if (msglen < sizeof(struct nvsp_message_header) +
1564		     sizeof(struct nvsp_4_send_vf_association)) {
1565		netdev_err(ndev, "nvsp_v4_msg length too small: %u\n", msglen);
1566		return;
1567	}
1568
1569	net_device_ctx->vf_alloc = nvmsg->msg.v4_msg.vf_assoc.allocated;
1570	net_device_ctx->vf_serial = nvmsg->msg.v4_msg.vf_assoc.serial;
1571
1572	if (net_device_ctx->vf_alloc)
1573		complete(&net_device_ctx->vf_add);
1574
1575	netdev_info(ndev, "VF slot %u %s\n",
1576		    net_device_ctx->vf_serial,
1577		    net_device_ctx->vf_alloc ? "added" : "removed");
1578}
1579
1580static void netvsc_receive_inband(struct net_device *ndev,
1581				  struct netvsc_device *nvscdev,
1582				  const struct vmpacket_descriptor *desc)
1583{
1584	const struct nvsp_message *nvmsg = hv_pkt_data(desc);
1585	u32 msglen = hv_pkt_datalen(desc);
1586
1587	/* Ensure packet is big enough to read header fields */
1588	if (msglen < sizeof(struct nvsp_message_header)) {
1589		netdev_err(ndev, "inband nvsp_message length too small: %u\n", msglen);
1590		return;
1591	}
1592
1593	switch (nvmsg->hdr.msg_type) {
1594	case NVSP_MSG5_TYPE_SEND_INDIRECTION_TABLE:
1595		netvsc_send_table(ndev, nvscdev, nvmsg, msglen);
1596		break;
1597
1598	case NVSP_MSG4_TYPE_SEND_VF_ASSOCIATION:
1599		if (hv_is_isolation_supported())
1600			netdev_err(ndev, "Ignore VF_ASSOCIATION msg from the host supporting isolation\n");
1601		else
1602			netvsc_send_vf(ndev, nvmsg, msglen);
1603		break;
1604	}
1605}
1606
1607static int netvsc_process_raw_pkt(struct hv_device *device,
1608				  struct netvsc_channel *nvchan,
1609				  struct netvsc_device *net_device,
1610				  struct net_device *ndev,
1611				  const struct vmpacket_descriptor *desc,
1612				  int budget)
1613{
1614	struct vmbus_channel *channel = nvchan->channel;
1615	const struct nvsp_message *nvmsg = hv_pkt_data(desc);
1616
1617	trace_nvsp_recv(ndev, channel, nvmsg);
1618
1619	switch (desc->type) {
1620	case VM_PKT_COMP:
1621		netvsc_send_completion(ndev, net_device, channel, desc, budget);
 
1622		break;
1623
1624	case VM_PKT_DATA_USING_XFER_PAGES:
1625		return netvsc_receive(ndev, net_device, nvchan, desc);
 
 
1626
1627	case VM_PKT_DATA_INBAND:
1628		netvsc_receive_inband(ndev, net_device, desc);
1629		break;
1630
1631	default:
1632		netdev_err(ndev, "unhandled packet type %d, tid %llx\n",
1633			   desc->type, desc->trans_id);
1634		break;
1635	}
1636
1637	return 0;
1638}
1639
1640static struct hv_device *netvsc_channel_to_device(struct vmbus_channel *channel)
1641{
1642	struct vmbus_channel *primary = channel->primary_channel;
1643
1644	return primary ? primary->device_obj : channel->device_obj;
1645}
1646
1647/* Network processing softirq
1648 * Process data in incoming ring buffer from host
1649 * Stops when ring is empty or budget is met or exceeded.
1650 */
1651int netvsc_poll(struct napi_struct *napi, int budget)
1652{
1653	struct netvsc_channel *nvchan
1654		= container_of(napi, struct netvsc_channel, napi);
1655	struct netvsc_device *net_device = nvchan->net_device;
1656	struct vmbus_channel *channel = nvchan->channel;
1657	struct hv_device *device = netvsc_channel_to_device(channel);
1658	struct net_device *ndev = hv_get_drvdata(device);
1659	int work_done = 0;
1660	int ret;
1661
1662	/* If starting a new interval */
1663	if (!nvchan->desc)
1664		nvchan->desc = hv_pkt_iter_first(channel);
1665
1666	nvchan->xdp_flush = false;
1667
1668	while (nvchan->desc && work_done < budget) {
1669		work_done += netvsc_process_raw_pkt(device, nvchan, net_device,
1670						    ndev, nvchan->desc, budget);
1671		nvchan->desc = hv_pkt_iter_next(channel, nvchan->desc);
1672	}
1673
1674	if (nvchan->xdp_flush)
1675		xdp_do_flush();
1676
1677	/* Send any pending receive completions */
1678	ret = send_recv_completions(ndev, net_device, nvchan);
1679
1680	/* If it did not exhaust NAPI budget this time
1681	 *  and not doing busy poll
1682	 * then re-enable host interrupts
1683	 *  and reschedule if ring is not empty
1684	 *   or sending receive completion failed.
1685	 */
1686	if (work_done < budget &&
 
1687	    napi_complete_done(napi, work_done) &&
1688	    (ret || hv_end_read(&channel->inbound)) &&
1689	    napi_schedule_prep(napi)) {
1690		hv_begin_read(&channel->inbound);
1691		__napi_schedule(napi);
1692	}
1693
1694	/* Driver may overshoot since multiple packets per descriptor */
1695	return min(work_done, budget);
1696}
1697
1698/* Call back when data is available in host ring buffer.
1699 * Processing is deferred until network softirq (NAPI)
1700 */
1701void netvsc_channel_cb(void *context)
1702{
1703	struct netvsc_channel *nvchan = context;
1704	struct vmbus_channel *channel = nvchan->channel;
1705	struct hv_ring_buffer_info *rbi = &channel->inbound;
1706
1707	/* preload first vmpacket descriptor */
1708	prefetch(hv_get_ring_buffer(rbi) + rbi->priv_read_index);
1709
1710	if (napi_schedule_prep(&nvchan->napi)) {
1711		/* disable interrupts from host */
1712		hv_begin_read(rbi);
1713
1714		__napi_schedule_irqoff(&nvchan->napi);
1715	}
1716}
1717
1718/*
1719 * netvsc_device_add - Callback when the device belonging to this
1720 * driver is added
1721 */
1722struct netvsc_device *netvsc_device_add(struct hv_device *device,
1723				const struct netvsc_device_info *device_info)
1724{
1725	int i, ret = 0;
1726	struct netvsc_device *net_device;
1727	struct net_device *ndev = hv_get_drvdata(device);
1728	struct net_device_context *net_device_ctx = netdev_priv(ndev);
1729
1730	net_device = alloc_net_device();
1731	if (!net_device)
1732		return ERR_PTR(-ENOMEM);
1733
1734	for (i = 0; i < VRSS_SEND_TAB_SIZE; i++)
1735		net_device_ctx->tx_table[i] = 0;
1736
1737	/* Because the device uses NAPI, all the interrupt batching and
1738	 * control is done via Net softirq, not the channel handling
1739	 */
1740	set_channel_read_mode(device->channel, HV_CALL_ISR);
1741
1742	/* If we're reopening the device we may have multiple queues, fill the
1743	 * chn_table with the default channel to use it before subchannels are
1744	 * opened.
1745	 * Initialize the channel state before we open;
1746	 * we can be interrupted as soon as we open the channel.
1747	 */
1748
1749	for (i = 0; i < VRSS_CHANNEL_MAX; i++) {
1750		struct netvsc_channel *nvchan = &net_device->chan_table[i];
1751
1752		nvchan->channel = device->channel;
1753		nvchan->net_device = net_device;
1754		u64_stats_init(&nvchan->tx_stats.syncp);
1755		u64_stats_init(&nvchan->rx_stats.syncp);
1756
1757		ret = xdp_rxq_info_reg(&nvchan->xdp_rxq, ndev, i, 0);
1758
1759		if (ret) {
1760			netdev_err(ndev, "xdp_rxq_info_reg fail: %d\n", ret);
1761			goto cleanup2;
1762		}
1763
1764		ret = xdp_rxq_info_reg_mem_model(&nvchan->xdp_rxq,
1765						 MEM_TYPE_PAGE_SHARED, NULL);
1766
1767		if (ret) {
1768			netdev_err(ndev, "xdp reg_mem_model fail: %d\n", ret);
1769			goto cleanup2;
1770		}
1771	}
1772
1773	/* Enable NAPI handler before init callbacks */
1774	netif_napi_add(ndev, &net_device->chan_table[0].napi, netvsc_poll);
 
1775
1776	/* Open the channel */
1777	device->channel->next_request_id_callback = vmbus_next_request_id;
1778	device->channel->request_addr_callback = vmbus_request_addr;
1779	device->channel->rqstor_size = netvsc_rqstor_size(netvsc_ring_bytes);
1780	device->channel->max_pkt_size = NETVSC_MAX_PKT_SIZE;
1781
1782	ret = vmbus_open(device->channel, netvsc_ring_bytes,
1783			 netvsc_ring_bytes,  NULL, 0,
1784			 netvsc_channel_cb, net_device->chan_table);
1785
1786	if (ret != 0) {
1787		netdev_err(ndev, "unable to open channel: %d\n", ret);
1788		goto cleanup;
1789	}
1790
1791	/* Channel is opened */
1792	netdev_dbg(ndev, "hv_netvsc channel opened successfully\n");
1793
1794	napi_enable(&net_device->chan_table[0].napi);
1795	netif_queue_set_napi(ndev, 0, NETDEV_QUEUE_TYPE_RX,
1796			     &net_device->chan_table[0].napi);
1797	netif_queue_set_napi(ndev, 0, NETDEV_QUEUE_TYPE_TX,
1798			     &net_device->chan_table[0].napi);
1799
1800	/* Connect with the NetVsp */
1801	ret = netvsc_connect_vsp(device, net_device, device_info);
1802	if (ret != 0) {
1803		netdev_err(ndev,
1804			"unable to connect to NetVSP - %d\n", ret);
1805		goto close;
1806	}
1807
1808	/* Writing nvdev pointer unlocks netvsc_send(), make sure chn_table is
1809	 * populated.
1810	 */
1811	rcu_assign_pointer(net_device_ctx->nvdev, net_device);
1812
1813	return net_device;
1814
1815close:
1816	RCU_INIT_POINTER(net_device_ctx->nvdev, NULL);
1817	netif_queue_set_napi(ndev, 0, NETDEV_QUEUE_TYPE_TX, NULL);
1818	netif_queue_set_napi(ndev, 0, NETDEV_QUEUE_TYPE_RX, NULL);
1819	napi_disable(&net_device->chan_table[0].napi);
1820
1821	/* Now, we can close the channel safely */
1822	vmbus_close(device->channel);
1823
1824cleanup:
1825	netif_napi_del(&net_device->chan_table[0].napi);
1826
1827cleanup2:
1828	free_netvsc_device(&net_device->rcu);
1829
1830	return ERR_PTR(ret);
1831}